Bank of America Corporation

Bank of America Corporaiton

Viju Kothuvatiparambil

Ramakrishna R Yannam

Donatus Asumu

Emad Noorizadeh

Priyank R Shah

David H Witting

Maruthi Z Shanmugam

Stephen Yeager

Mahalakshmi Chandra Sekaran

Matthew Hsieh

Sushil Golani

Bharathiraja Krishnamoorthy

Karen Trevino

Marco Ramirez

Luis Gerardo Mojica De La Vega

Michael Bostian

Jamaal C Long

A system for determining intents associated with human utterances at a voice response system is provided. The system may involve a machine learning classifier and a deep learning classifier. The machine learning classifier may receive labeled training data. The labeled training data may be sample utterances. Each sample utterance may be assigned an intent. The intent may be the label of the utterance. The intent may be included in a predefined plurality of intents. The machine learning classifier may include a feature engineering and extraction module. The feature engineering and extraction module may identify features specific to each intent. The machine learning classifier may label previously unlabeled utterances received in a production environment. The newly labeled utterances received in the production environment may be used as training data to train a deep learning classifier. After being trained, the deep learning classifier may receive and label previously unlabeled utterances.

Deep learning system

Systems and methods for automated, degradation-resistant, tuning of machine-learning ('ML') models are provided. The systems and methods may identify an inoperative input utterance, retrieve a feature set associated with the inoperative input utterance, and generate an updated utterance-feature-intent (“UFI”) mapping based on the retrieved feature set. The systems and methods may retrain the ML model using the updated UFI mapping, and compare the accuracy of the system after the retraining and before the retraining. In a scenario where the accuracy of the system does not improve, the systems and methods may amplify the updated UFI mapping. In a scenario where the accuracy of the system does improve, the systems and methods may deploy the updated UFI mapping.

Systems and methods for automated degradation-resistant tuning of machine-learning language processing models

Aspects of the disclosure relate to supervised machine-learning ('ML') training platforms for artificial intelligence ('AI') computer systems. The ML training platform may include isolated update testing. The isolated update testing may feature a plurality of environments with various levels of isolation. The ML training platform may also include bi-directional channels for controlled update propagation.

Supervised machine-learning training platform with isolated update testing and bi-directional update propogation

Apparatus for mapping a user utterance onto a plurality of intents is provided. The apparatus may include an intent training database that includes a plurality of tokens and intents. The apparatus may include a processor. The processor may utilize a token-intent map to generate a token-row map and an intent-column map. The processor may map the plurality of tokens onto a token-intent matrix. The processor may generate a token-cognitive matrix, a cognitive-comprehension matrix and an intent-cognitive matrix from the decomposition. The cognitive-comprehension matrix may be the space of entanglement between the token-cognitive matrix and the intent-cognitive matrix. The processor may reduce the rank of the cognitive-comprehension matrix. The processor may compute a plurality of token vectors from a computation of the token-cognitive matrix and the cognitive-comprehension matrix. The processor may compute a plurality of intent-cognitive vectors from a computation of the cognitive-comprehension matrix and the intent-cognitive matrix.

Quantum technology for use with extracting intents from linguistics

A system for extracting verifiable entities from a user-utterance received on an automated calling service is provided. The system may include a receiver configured to receive a user-utterance, a processor and a non-transitory computer-readable media comprising computer-executable instructions. The processor may be configured to execute the instructions which, canonicalize the user-utterance into a plurality of tokens, determine the number of tokens of the user-utterance, and generate, using a sliding-window protocol, a comprehensive number of n-gram sequences from the user-utterance. The processor may be configured to process a plurality of threads of execution that may include a series of actions executed on the n-gram sequences to identify and extract verified entities from the user-utterance.

Deterministic multi-length sliding window protocol for contiguous string entity

A language processor includes a parser and an invoker. The parser parses a spoken statement to detect a plurality of words in the spoken statement and generates a parse tree based on the detected plurality of words. The invoker determines, based on the parse tree, a plurality of potential services to invoke to respond to the spoken statement, calculates a confidence score for each potential service of the plurality of potential services, wherein each confidence score indicates a likelihood, that a potential service is a service requested by the spoken statement, and determines, based on the calculated confidence scores, to invoke a potential service of the plurality of potential services in response to the spoken statement. This helps reduce a number of questions that a caller associated with the spoken statement has to answer before the language processor determines a service to invoke to respond to the caller.

Automated response tool to reduce required caller questions for invoking proper service

Methods for analyzing a query is provided. Methods may include receiving a query from a user. Methods may include dissecting the query into a plurality of word segments. Methods may include determining, based on a predetermined ontology, which of the plurality of words satisfies a predetermined ontological threshold and which words do not satisfy the predetermined ontological threshold. Methods may include discarding the words that do not satisfy the predetermined ontological threshold. Methods may include determining a part of speech associated with each word based on the content of the query. Methods may include determining a concept associated with each word based on the content of the query. Methods may include displaying to the user word that satisfies the predetermined threshold alongside the associated part of speech and associated concept. Methods may include enabling the user to change each word, each concept and each part of speech.

Query analyzer

A method for utilizing a multi-intent matrix is provided. The method may be used to avoid a derived intent collision while determining correct intent of a human-voice-telephone query. The method may include receiving a telephone call at the voice response system. The method may include responding to the telephone call at the voice response system. The response may include prompting a human caller to utter a reason for the telephone call. The human caller may be associated with the telephone call. The method may include receiving a human-voice-telephone-query-utterance at the voice response system. The method may include translating the human-voice-telephone-query-utterance into a text string. The translating may occur at the voice response system or a translation module. The method may include analyzing the text string utilizing a group of parameters. The method may include re-analyzing the text string utilizing a group of parameters set to custom values.

Derived intent collision detection for use in a multi-intent matrix

An automated response tool includes a receiver, a language processor, and a service invoker. The receiver receives a call and a spoken statement from the received call. The language processor detects a plurality of words in the spoken statement and generates a parse tree based on the detected plurality of words. The language processor also determines, based on the parse tree, a service to invoke in response to the spoken statement, determines a parameter used during execution of the determined service, and determines, based on the parse tree, a value of the determined parameter. The service invoker issues a command to invoke the determined service, wherein the command comprises the determined value of the determined parameter.

Automated response tool

A language processor includes a parser, an invoker, and an extractor. The parser parses a spoken statement to detect a plurality of words in the spoken statement and generates a parse tree based on the detected plurality of words. The invoker determines, based on the parse tree, a plurality of potential services to invoke to respond to the spoken statement. The extractor determines, for each potential service of the plurality of potential services, a parameter used during execution of that potential service and a value of the determined parameter based on the parse tree. The invoker is further configured to issue a command to invoke a potential service of the plurality of potential services using a value of a determined parameter.

An apparatus for automatic completion of an entry into an input field is provided. The apparatus may include a webpage. The webpage may include a search box. The search box may include a user input field. The apparatus may further include a processor and a machine-readable memory. When a user begins to type a query into the user input box, the processor may be configured to complete, pending user confirmation, the user's query. The completion may be based, for example, on information stored on the machine-readable memory pertaining to the user. The information may be available to an authorized accessor of the machine-readable memory via a secure communications channel. However, the information may be restricted from the user's access via a communications channel associated with the user.

Method and apparatus for automatic completion of an entry into an input field

A method for natural language search for variables is provided. The method may include searching an index using key words from a user's natural language question and the context of the user's question. The index may reference variables and/or web service calls in a domain model. The method may also include saving documents obtained in response to the search. The method may also include mapping each of the documents as a node into an object graph. Each node may be associated with a parent node, except when the node is a root node. The method may also include identifying the root node of each document. The method may also include identifying the path of each node from the node to the node's root node. The method may also include identifying matching paths. Each matching path may provide an answer to the user's question.

Method and apparatus for natural language search for variables

Methods for extracting queries from webpages is provided. Methods may include receiving a plurality of queries. Each query in the plurality may be input into a search box located on a public-facing webpage. Methods may include receiving content associated with each of the public-facing webpages. Methods may include receiving identifying information associated with an inputter of each query. Methods may include selecting at least one query from the plurality of queries based at least in part on factors. The factors may include query length. The factors may include query rank. The query rank may be based in part on the magnitude of predetermined terminology included in the at least one query. The predetermined terminology may be stored on a computer-readable memory. The factors may include the grammatical relationship between query terms. The factors may include the identifying information about the inputter associated with the at least one query.

Growing community of inventors

Seattle, WA, United States of America

Viju Kothuvatiparambil